AI curriculum
Course outline for Retrieval-Augmented Generation
Build a question-answering application that retrieves relevant document passages before generating an answer. Prepare documents, create embeddings and an index, compare retrieval methods, and construct useful model context. Generate cited answers and investigate failures by separating retrieval performance from answer support. Compare graph-assisted retrieval without making a graph database the starting requirement.
Retrieval-Augmented Generation Course Objectives
- Explain when RAG helps and how it differs from fine-tuning or direct context.
- Ingest documents, create traceable chunks, and build an embedding index.
- Compare lexical, vector, hybrid, and reranked retrieval on fixed questions.
- Generate answers using selected passages and check their citations.
- Measure retrieval and answer support separately and diagnose pipeline failures.
- Explain when relationship-based retrieval is useful and when it adds unnecessary complexity.
Pre-requisites
- Basic programming, file processing, and HTTP/JSON use.
- Applied Generative AI or equivalent ability to submit model requests and inspect responses.
- Artificial Intelligence Foundations or equivalent knowledge is expected; How Large Language Models Work and Building Applications with LLMs are useful preparation, not mandatory completions.
- No vector database, graph database, or advanced mathematics is assumed.
Before this course
Complete the required courses or bring equivalent experience. Recommended courses add useful depth; they are not enrolment gates.
Required preparation
Recommended preparation
Readiness checks
- Basic programming, file processing, and HTTP/JSON use.
- Applied Generative AI or equivalent ability to submit model requests and inspect responses.
- Artificial Intelligence Foundations or equivalent knowledge is expected; How Large Language Models Work and Building Applications with LLMs are useful preparation, not mandatory completions.
- No vector database, graph database, or advanced mathematics is assumed.
Can program, process files, use HTTP/JSON, and submit and inspect model requests; no vector database, graph database, or advanced mathematics is assumed.
Find your place in the learning mapLab Setup
Delivery requirements: the following environment, materials, access, and exercise timing must be prepared and verified before delivery. This outline does not certify that a lab package is ready.
A selected programming runtime, a local lexical/vector retrieval library, embedding and reranking models, a small source corpus, and a generation endpoint. Preload larger model files before class. A hosted vector service is not required. Learners generate embeddings and answers during the lab. The instructor must verify resource needs and inference access; cached outputs support debugging only.
Detailed Course Outline
Module 1: From documents to a searchable collection (2 hours)
- RAG and other context approaches: Model knowledge, supplied context, RAG, and fine-tuning roles.
- Document ingestion: Load and clean documents while preserving source identifiers and useful metadata.
- Traceable chunks: Chunk boundaries, size, overlap, and preserved source context.
- Compatible embeddings: Choose embedding models with compatible query/document representations and meaningful similarity comparisons.
- Local vector indexes: Building and updating a small local vector index.
Practical outcome: Ingest a small corpus and inspect actual chunks, vectors, and source references.
Module 2: Finding and ranking relevant passages (2 hours)
- Lexical retrieval: Run lexical retrieval and inspect passages selected through exact-term matches.
- Vector retrieval: Use vector similarity, top-k selection, and relevance judgments on fixed questions.
- Metadata filtering: Apply metadata filters to control the data scope at query time.
- Hybrid retrieval: Compare hybrid candidate retrieval and the choices used to combine scores.
- Reranking: Apply reranking and compare the result with the initial retrieval order.
Practical outcome: Run retrieval variants on the same questions and compare the evidence found.
Module 3: From passages to grounded answers (2 hours)
- Context construction: Construct context through passage selection, ordering, limits, and duplicate removal.
- Answer and source instructions: Specify answer instructions and source-reference handling before the generation step.
- Grounded generation: Run actual model generation from retrieved passages in the question-answering pipeline.
- Claim-level citation checks: Citation presence versus support for a specific claim.
- Problematic retrieved evidence: Handle missing, conflicting, stale, and untrusted retrieved content without inventing evidence.
Practical outcome: Produce a cited answer and an appropriate insufficient-evidence response.
Module 4: Evaluation, diagnosis, and graph-assisted retrieval (2 hours)
- Retrieval metrics: Labeled questions and retrieval precision/recall at a fixed cutoff.
- Answer and citation assessment: Answer correctness, relevance, supported claims, and citation checks.
- Configuration comparisons: Compare a baseline with one changed chunking or retrieval configuration.
- Pipeline failure diagnosis: Ingestion, retrieval, ranking, context, and generation failure diagnosis.
- Graph-assisted retrieval comparison: Graph RAG: entity relationships, a worked traversal, and added operating cost.
Practical outcome: Measure a pipeline change and justify whether graph assistance is needed.
Practical exercise
- Build a small handbook question-answering pipeline from document loading through actual answer generation. Generate corpus embeddings, index them locally, and inspect returned passages. Run lexical, vector, and hybrid searches, then apply a reranker to selected queries. Compare a baseline and one revised configuration on separate development and held-out questions. Include a question with no supporting document and check claimed citations against their passages.
What you will produce
Build a corpus-to-answer RAG pipeline, compare retrieval methods, and measure retrieval and source-supported answer quality separately.
How the work is checked
- New document text is converted into searchable, source-identifiable chunks.
- The answer uses retrieved context from a running pipeline, not a supplied answer table.
- Search comparisons hold the questions and judgments fixed.
- Retrieval measures are separate from answer and citation assessments.
- The learner locates a failure at a specific pipeline stage.
- The graph comparison identifies the useful relationship and a simpler alternative.
Eight guided hours include demonstrations, practice, feedback, and questions. Breaks and bulk downloads are outside the total. Core practice includes actual embeddings, retrieval, reranking, and answer generation. Graph RAG is a sourced worked comparison, not a second graph-store implementation. Large-scale indexing, permission-system implementation, ontology design, and fine-tuning are outside scope. Preload larger model files and verify resources, generation access, and the selected local libraries before delivery. Cached outputs support debugging only and do not replace the required live pipeline runs.
When another approach is better
- Use direct context when a small, stable document set fits without a retrieval pipeline.
- Do not add a graph store when lexical or vector retrieval meets the relationship needs.
- Use specialist infrastructure training for large-scale indexing and permission enforcement.
How we train
Contact us for full course details, including duration, delivery options and lab requirements.
Ask about exercises, instructor feedback and the prior knowledge you need. Tool-specific courses marked provisional may become modules in a broader course.
Discuss your learning goalsRelated Courses
These courses list this course as preparation.